Difficulty-to-measure　process　parameter　relative　to　production　quality　and　efficient　of　complex　industrial　process　is　obtained　normally　by　off-line　analysis　or　expert　estimation.　One　of　the　main　reason　is　that　the　soft　measuring　model　between　multi-source　input　features　and　such　process　parameter　is　difficulty　to　be　constructed.　Aim　at　the　above　issue,　a　new　soft　measuring　method　is　proposed　in　this　study.　At　first,　linear　and　nonlinear　feature　sub-sets　are　selected　by　using　correlation　coefficient　and　mutual　information　method.　Then,　four　types　of　linear　and　nonlinear　candidate　sub-models　are　constructed　based　on　the　above　feature　subsets.　At　last,　optimization　and　weighting　algorithms　are　used　to　select　and　combine　the　selected　ensemble　sub-models.　Thus,　the　final　selective　ensemble　learning-based　soft　measuring　model　is　obtained.　The　modeling　results　based　on　high　dimensional　mechanical　vibration　frequency　spectrum　validate　effectiveness　of　this　approach.　©　2019　IEEE.